Synthesizing images of clothing on models
Neural networks of suitable topology are trained with pairs of images, where one image of each pair depicts a garment, and the other image of each pair depicts the garment being worn by a model. Once trained, the neural network can synthesize an image based on a new image of a garment, where the synthesized image could plausibly have appeared in the training set, paired with the new image of the garment. Quantitative parameters controlling the image synthesis permit adjustment of features of the synthetic image, including the skin tone, body shape and pose of the model, accessories depicted in the synthetic image, and characteristics of the garment as depicted, such as length, sleeve style, collar style or tightness.
1. A method comprising:
initializing a Generative Adversarial Network (“GAN”) comprising an analysis neural network and a synthesis neural network;
training the analysis neural network to recognize characteristics of an input image of a training garment alone and encode the characteristics into a training Z-Vector having fewer components than the input image of the training garment alone has pixels, said training producing a trained analysis neural network;
training the synthesis neural network to create a training synthetic image resembling a model wearing the training garment from the training Z-Vector, said training producing a trained synthesis neural network;
presenting an instance image of an instance garment alone to the trained analysis neural network and obtaining a corresponding instance Z-Vector from the trained analysis neural network;
presenting the instance Z-Vector to the trained synthesis neural network and obtaining a corresponding instance synthetic image from the trained synthesis neural network, said instance synthetic image resembling a model wearing the instance garment; and
displaying the instance synthetic image.
2. The method of claim 1 , further comprising:
determining an effect of a subset of the training Z-Vector on a recognizable characteristic of the training synthetic image;
adjusting a subset of the instance Z-Vector to produce a modified instance Z-Vector; and
presenting the modified instance Z-Vector to the trained synthesis neural network and obtaining a corresponding modified synthetic image from the trained image synthesis neural network, said modified synthetic image being similar to the instance synthetic image, but different in the recognizable characteristic; and
displaying the modified synthetic image.
3. The method of claim 2 wherein the recognizable characteristic of the training synthetic image is one of model skin tone, model pose, model body weight, model body shape, accessories, garment length, garment sleeve style, garment collar style or garment fit tightness.
4. The method of claim 1 wherein the synthesis neural network also receives information from at least one of:
a node of the analysis neural network at a similar depth to a receiving node of the synthesis neural network; or
a node of the analysis neural network at a different depth from the receiving node of the synthesis network.
5. A method comprising:
initializing a multi-layer neural network comprising an analysis neural network and a synthesis neural network;
training the analysis neural network to recognize characteristics of an input image of a garment alone and encode the characteristics into a Z-Vector having fewer components than the input image of the garment alone has pixels, said training producing a trained analysis neural network;
training the synthesis neural network to create an image resembling a model wearing the garment from the Z-Vector, said training producing a trained synthesis neural network;
presenting an instance image of an instance garment alone to the trained analysis neural network and obtaining a corresponding instance Z-Vector from the trained analysis neural network;
presenting the instance Z-Vector to the trained synthesis neural network and obtaining a corresponding synthetic image from the trained synthesis neural network, said synthetic image resembling a model wearing the instance garment; and
displaying the synthetic image.
6. The method of claim 5 wherein the multi-layer neural network is one of a Generative Adversarial Network, a Recurrent Neural Network, a Recurrent Inference Machine, or a Variational Autoencoder.